[Python] Why is the output of id() applied to a Python object returned from C++ via pybind11...

I am using a holder class to store a std::shared_ptr to another C++ class created in Python thus:

#include <memory>
#include <iostream>

#include <pybind11/pybind11.h>

class Child {

};

class Holder {
public:
Holder(std::shared_ptr<Child> child) : m_child {child} {
print_child_address();
}

std::shared_ptr<Child> get_child() {
return m_child;
}

void print_child_address() {
std::cout << m_child << std::endl;
}

private:
std::shared_ptr<Child> m_child;
};

namespace py = pybind11;

PYBIND11_MODULE(pointers, m) {

py::class_<Child, std::shared_ptr<Child>>(m, "Child")
.def(py::init<>());

py::class_<Holder>(m, "Holder")
.def(py::init<const std::shared_ptr<Child>&>(), py::arg("child"))
.def("get_child", &Holder::get_child)
.def("print_child_address", &Holder:rint_child_address);
}


CMakeLists.txt:

cmake_minimum_required(VERSION 3.21)
project(pointers VERSION 1.0 LANGUAGES CXX)

list(APPEND CMAKE_PREFIX_PATH "${CMAKE_BINARY_DIR}")
find_package(pybind11 REQUIRED)

pybind11_add_module(pointers pointers.cpp)


This works from Python:

>>> from pointers import *
>>> c = Child()
>>> h = Holder(c)
0x607d93adc800
>>> hex(id(c))
'0x7c9f96460a70'
>>> hex(id(h.get_child()))
'0x7c9f96460a70'


However, when I call it like this, I get a very strange result:

>>> h = Holder(Child())
0x6304fb8a8800
>>> hex(id(h.get_child()))
'0x7dd87bcc4eb0'
>>> hex(id(h.get_child()))
'0x7dd87bcc4c70'
>>> a = h.get_child()
>>> hex(id(h.get_child()))
'0x7dd87bcc4eb0'
>>> hex(id(h.get_child()))
'0x7dd87bcc4eb0'


The id (address) of h.get_child() changes every time until I assign h.get_child() to a variable at which point it becomes stable. In iPython the behaviour is even stranger, in that the address becomes stable after calling h.get_child() on its own, without an assignment.

(Co-incidentally, in this run the final stable address is the same address that I received at the beginning, however this is not normally the case.)

The std::shared_ptr inside the Holder class is stable and everything seems to work on the C++ side but I don't like this apparent instability of the Python object. Can anyone explain what is happening here?

Continue reading...